Skip to main contentThe Local Bench

Use case 6 of 13

Smarter camera alerts

Get "a courier left a parcel at the door" instead of "motion detected", with the video never leaving your network.

A small white home security camera wired with a blue cable to a black mini PC
Image: generated with Grok Imagine

The real example

Frigate NVR with generative AI descriptions

Frigate is an open-source network video recorder with real-time object detection. Version 0.18 can send thumbnails of tracked objects to a vision-language model to write a description, used for semantic search and review summaries. It supports local providers such as Ollama and llama.cpp, and recommends Qwen3-VL for descriptions. Its docs say most 7B 4-bit vision models fit in 8 GB of VRAM and recommend an NVIDIA GPU or Apple silicon for the model host.

Sources: Frigate: generative AI object descriptions · Frigate: configuring generative AI

Why run it locally

Camera footage of your home, your neighbours and the street should not stream to someone else's server. Local detection also keeps working offline.

Recommended stack

Three tiers built from our builds. Weights-in-memory is our arithmetic (parameters × bits), checked against each device's usable memory; it excludes KV cache and runtime overhead. Runtime compatibility is not validated.

Budget

Silent home server

$1,283Sample

16 GB on GeForce RTX 4060 Ti 16 GB

Weights ≈ 5.0 GB of 15.5 GB usableEstimate

Excludes KV cache and runtime overhead. Runtime compatibility is not validated.

A 7B VLM describes events after the fact; detection itself runs on a small detector.

Runtime: Frigate + Ollama

Setup outline

  1. Install Frigate with Docker and add your RTSP cameras with a detector (GPU, Coral or OpenVINO).
  2. Run Ollama or llama.cpp on the GPU host and pull a vision model.
  3. Add a genai provider in Frigate pointing at the local server, with the descriptions role.
  4. Turn on semantic search and test queries such as "person carrying a box".
  5. Route alerts through Home Assistant notifications.

Limits

  • Descriptions arrive seconds after the event. Use plain object detection for anything time-critical.
  • VLMs can misdescribe. Never automate locks or alarms on a generated sentence alone.
  • Check local laws on recording public areas and neighbours' property.

Next: A medical literature helper (not medical advice)

Build prices come from our build datasheets: GPUs and devices are sourced retail or sold prices; other parts are sample prices until sourced, and the flag on each total says which. Model links go to our datasheets, where each parameter count and licence was checked against the Hugging Face API.